MATHEMATICS 3103 (Functional Analysis)

YEAR 20122013, TERM 2

HANDOUT #7: THE BAIRE CATEGORY THEOREM AND ITS
CONSEQUENCES
We shall begin this last section of the course by returning to the study of general metric
spaces, and proving a fairly deep result called the Baire category theorem.1 We shall then
apply the Baire category theorem to prove three fundamental results in functional analysis:
the Uniform Boundedness Theorem, the Open Mapping Theorem, and the Closed Graph
Theorem.

The Baire category theorem

Let X be a metric space. A subset A X is called nowhere dense in X if the interior of
the closure of A is empty, i.e. (A) = . Otherwise put, A is nowhere dense iff it is contained
in a closed set with empty interior. Passing to complements, we can say equivalently that A
is nowhere dense iff its complement contains a dense open set (why?).
Proposition 7.1 Let X be a metric space. Then:
(a) Any subset of a nowhere dense set is nowhere dense.
(b) The union of finitely many nowhere dense sets is nowhere dense.
(c) The closure of a nowhere dense set is nowhere dense.
(d) If X has no isolated points, then every finite set is nowhere dense.
Proof. (a) and (c) are obvious from the definition and the elementary properties of closure
and interior.
To prove (b), it suffices to consider a pair of nowhere dense sets A1 and A2 , and prove
that their union is nowhere dense (why?). It is also convenient to pass to complements,
and prove that the intersection of two dense open sets V1 and V2 is dense and open (why is
this equivalent?). It is trivial that V1 V2 is open, so let us prove that it is dense. Now,
a subset is dense iff every nonempty open set intersects it. So fix any nonempty open set
U X. Then U1 = U V1 is open and nonempty (why?). And by the same reasoning,
U2 = U1 V2 = U (V1 V2 ) is open and nonempty as well. Since U was an arbitrary
nonempty open set, we have proven that V1 V2 is dense.
To prove (d), it suffices to note that a one-point set {x} is open if and only if x is an
isolated point of X; then use (b). 
Proved (for Rn ) by the French mathematician Rene-Louis Baire (18741932) in his 1899 doctoral thesis.
Baire made a number of important contributions to real analysis in addition to the category theorem.
However, it turns out that the Baire category theorem for the real line was actually proved two years earlier,
in 1897, by the American mathematician William Fogg Osgood (18641943)!
1

(a) and (b) can be summarized by saying that the nowhere dense sets form an ideal of
sets.
Example. The Cantor ternary set C consists of all real numbers in the interval [0, 1]
that can be written as a ternary (base-3) expansion in which the digit 1 does not occur, i.e.

P
x=
an /3n with an {0, 2} for all n. Equivalently, C can be constructed from [0, 1] by
n=1

deleting the open middle third of the interval [0, 1], then deleting the open middle thirds of
each of the intervals [0, 1/3] and [2/3, 1], and so forth. If Cn denotes the union of the 2n

T
closed intervals of length 1/3n that remain at the nth stage, then C =
Cn . It follows
n=1

immediately that C is closed (and indeed compact). Moreover, since Cn contains no open
interval of length greater than 1/3n , it follows that C contains no open interval at all, i.e. C
has empty interior and hence is nowhere dense. 
Although the union of finitely many nowhere dense sets is nowhere dense, the union of
countably many nowhere dense sets need not be nowhere dense: for instance, in X = R, the
rationals Q are the union of countably many nowhere dense sets (why?), but the rationals
are certainly not nowhere dense (indeed, they are everywhere dense, i.e. (Q) = Q = R).
This observation motivates the introduction of a larger class of sets: A subset A X
is called meager (or of first category) in X if it can be written as a countable union of
nowhere dense sets. Any set that is not meager is said to be nonmeager (or of second
category). The complement of a meager set is called residual.
We then have as an immediate consequence:
Proposition 7.2 Let X be a metric space. Then:
(a) Any subset of a meager set is meager.
(b) The union of countably many meager sets is meager.
(c) If X has no isolated points, then every countable set is meager.
(a) and (b) can be summarized by saying that the meager sets form a -ideal of sets. In
a certain topological sense, the meager sets can be considered small and even negligible.
Warnings. 1. Be careful of the terminology, which can be confusing. The meager sets
are in some sense small. The residual sets are in some sense large (i.e. their complements
are small). But the second category sets are not necessarily large; they are merely not
small.
2. Note also that meager, nonmeager and residual are attributes not of a set A
in and of itself, but of a set A in the metric space X. Which category a set has depends on
the space within which it is considered. For example, a line is residual (and, we will soon
show, nonmeager) inside itself, but it is nowhere dense (and hence meager) inside a plane.
Similarly, Z is residual and nonmeager inside itself indeed, in Z every set is open (why?),
so the only meager set is (why?) but Z is nowhere dense (and hence meager) inside R.

2

Remark. If you have studied Measure Theory, then you have encountered another
important -ideal of sets in R or Rn , namely the sets of Lebesgue measure zero (also
called null sets). These sets are negligible in the measure-theoretic sense.
The meager sets and the measure-zero sets thus constitute two -ideals, each of which
properly contains the -ideal of countable sets.2 It is natural to ask whether these properties
are related. For instance, does one of the two classes contain the other? It turns out that the
answer is no, and that the two notions of smallness can in some cases even be diametrically
opposed. In fact, it is not difficult to prove that the real line can be decomposed into two
complementary sets, one of which is meager and the other of which has measure zero. So
a set that is small in one of these senses can be very big in the other sense, and vice
versa. There is of course nothing intrinsically paradoxical about the fact that a set that is
small in one sense may be large in another sense.
But despite the fact that there is no necessary relation between the properties of being
meager and measure zero, it turns out that there is an striking analogy between these properties. Many (but not all) theorems about meager sets have analogues (albeit sometimes
with quite different proofs) for sets of measure zero, and conversely. A beautiful (and very
readable) book about this analogy is John C. Oxtoby, Measure and Category. 
We are now ready to state the Baire category theorem:
Theorem 7.3 (Baire category theorem) Let X be a complete metric space. Then:
(a) A meager set has empty interior.
(b) The complement of a meager set is dense. (That is, a residual set is dense.)
(c) A countable intersection of dense open sets is dense.
You should carefully verify that (a), (b) and (c) are equivalent statements, obtained by
taking complements.
In applications we frequently need only the weak form of the Baire category theorem that
is obtained by weakening is dense in (b,c) to is nonempty (which is valid whenever X
is itself nonempty):
Corollary 7.4 (weak form of the Baire category theorem) Let X be a nonempty complete metric space. Then:
(b) X cannot be written as a countable union of nowhere dense sets. (In other words, X
is nonmeager in itself.)
(b ) If X is written as a countable union of closed sets, then at least one of those closed
sets has nonempty interior.
(c) A countable intersection of dense open sets is nonempty.
2

To see that the containment is proper, it suces to observe that the Cantor ternary set is meager,
measure-zero (why?) and uncountable (why?).

As preparation for the proof of the Baire category theorem, let us prove a useful lemma
due to Georg Cantor.3 Recall first that if X is a compact metric space, then any decreasing
sequence F1 F2 F3 . . . of nonempty closed sets has a nonempty intersection (why?).
This is not true in general in a noncompact metric space (even one that is complete and
separable): for instance, in X = R (or N) consider Fn = [n, ). But if we add the additional
hypothesis that the diameters of the sets Fn tend to zero, then it is true, provided only that
X is complete:
Lemma 7.5 (Cantor) Let X be a complete metric space, and let F1 F2 F3 . . . be a
decreasing sequence of nonempty closed subsets of X, with diam Fn 0. Then there exists

T
T
a point x X such that
Fn = {x}. In particular,
Fn 6= .
n=1

n=1

Proof. In each set Fn choose a point xn . Then the sequence (xn ) is Cauchy: for if m, n N
we have d(xm , xn ) diam FN (why?), which tends to zero as N . Since X is complete,
the sequence (xn ) has a limit x. But since xn FN for all n N, and FN is closed, we
T


T
have x FN . Since this holds for all N, we have x
FN . But since diam
FN
inf diam(FN ) = 0, we must have

N 1

n=1

N =1

N =1

Fn = {x}. 

Proof of the Baire category theorem. It is easily seen that (a) and (b) are
equivalent statements. Now, if A1 , A2 , . . . is a sequence of nowhere dense sets, then A1 , A2 , . . .
is a sequence of nowhere dense closed sets (why?) satisfying An An ; so to prove (b)
it clearly suffices to prove that a countable union of nowhere dense closed sets has a dense
complement. But this is equivalent, by complementation, to the assertion that a countable
intersection of dense open sets is dense, i.e. statement (c). So this is what we shall prove.
The proof follows the same idea as in the proof of Proposition 7.1, but with infinitely many
steps rather than just two, and with some minor modifications to allow us to exploit Cantors
lemma. Let V1 , V2 , . . . be the given sequence of dense open sets, and fix any nonempty open
set U X. Fix also a sequence (an ) of positive numbers tending to zero (say, an = 1/n).
The set U V1 is open and nonempty (why?), so we can choose a closed ball B(x1 , r1 ) inside
it, with radius 0 < r1 < a1 . Then B(x1 , r1 ) V2 is open and nonempty (why?), so we can
choose a closed ball B(x2 , r2 ) inside it, with radius 0 < r2 < a2 . And so forth: at the nth
stage we observe that B(xn1 , rn1 ) Vn is open and nonempty, so we can choose a closed
ball B(xn , rn ) inside it, with radius 0 < rn < an . Then Cantors lemma tells us that the
sequence of closed balls B(xn , rn ) has nonempty intersection (why?). But B(xn , rn ) U Vn
T


(why?), so it follows that U

Vn is nonempty. Since U was an arbitrary nonempty
n=1

open set, we have proven that

Vn is dense.4 

n=1
3

German mathematician Georg Cantor (18451918) revolutionized mathematics by founding, almost

single-handedly, the modern theory of innite sets. He also made important contributions to real analysis
and number theory.
4

If you worry about set-theoretic questions, you may have observed that some form of the axiom of
choice was invoked implicitly in this proof, in order to choose the closed ball B(xn , rn ). It turns out that

Example. We have seen that the rationals Q form a meager subset of the real line. What
about the irrationals R \ Q? Are they meager? The answer is not immediately obvious from
the definition of meager (we would have to consider all possible ways of writing R \ Q as
a countable union). But the answer follows immediately from the Baire category theorem:
for if R \ Q were meager, then so would be Q (R \ Q) = R, contradicting the (weak form
of the) Baire category theorem. 
Remark. When I said at the beginning of these notes that the Baire category theorem
is a fairly deep result, I did not mean that its proof is especially difficult as you have
seen, it is not (the proof of the equivalence of the three notions of compactness was more
difficult in my opinion). But what is deep is the mere idea to consider the countable unions
of nowhere dense sets. This was a stroke of genius on Baires (and Osgoods) part, and it
has had enormously powerful consequences in both real analysis and functional analysis
of which we will see merely a few examples.

Some applications of the Baire category theorem

Before beginning the applications to functional analysis, I would like to give just a little
bit of the flavor of the results that can be obtained in real analysis using the Baire category
theorem.
Example 1: F and G sets. Let X be a metric space. A subset A X is called an
F set if it can be written as a countable union of closed sets. It is called a G set if it
can be written as a countable intersection of open sets. Clearly, a set is F if and only if its
complement is G . Also, in a metric space (though not in a general topological space) every

T
closed set is a G : it suffices to observe that if A is closed, then A =
{x: d(x, A) < 1/n}
n=1

(why?) and that the latter sets are open (why?). By the same token, every open set is an
F .
Now let X = R. The rationals Q are clearly an F (why?) and hence the irrationals R \ Q
are a G . What about the reverse? Are the rationals also a G (and hence the irrationals
also an F )?
It is not so easy to answer these questions by elementary methods. But using the Baire
category theorem one can give an easy answer no. We know that the rationals are an F ;
what was needed is a weak form of the axiom of choice called the axiom of dependent choices (see e.g.
http://en.wikipedia.org/wiki/Baire category theorem). Indeed, logicians have proven that the axiom
of dependent choices is equivalent in ZermeloFraenkel (ZF) set theory to the Baire category theorem.
On the other hand, the Baire category theorem for separable metric spaces can be proven in ZF set theory
without any choice axiom: it suces to x a dense sequence (yn ) in X and an enumeration (qn ) of the
positive rational numbers, and then choose (xn , rn ) to be the first (in lexicographic order) pair (yk , qk ) that
has the desired property.
So the oft-made statement that the Baire category theorem provides nonconstructive existence proofs is
not quite right: when applied to separable metric spaces (which are indeed the principal applications), the
Baire category method provides in principle a construction by successive approximation of the elements of
the space that are claimed to exist. But this construction may be too complicated to be useful, so for all
practical purposes the method may be treated as nonconstructive.

so if also the irrationals were an F , then the whole real line could be written as a countable
union of closed sets, each one of which is entirely contained either in Q or in R \ Q. But
Corollary 7.4b to the Baire category theorem tells us that one of these closed sets would
have to have nonempty interior, which contradicts the fact that it is contained in Q or R \ Q.
We can also make the following trivial but useful observation:
Proposition 7.6 Let X be a metric space. Then an F subset A X is either meager or
else has nonempty interior.
Proof. By hypothesis A can be written as a countable union of closed sets. If all of these
closed sets have empty interior, then A is meager; if at least one of them has nonempty
interior, then so does A. 
Remark. Proposition 7.6 holds whether or not X is complete (all we used was the
definition of meager). But if X is not complete, then a meager set need not be small;
indeed, it could be all of X! (E.g. if X = Q considered with the metric inherited from R.)
In particular, when X is not complete, the two alternatives given in Proposition 7.6 need
not be mutually exclusive: a set could be meager and have nonempty interior. On the other
hand, if X is complete, then the meager sets really are small, i.e. they have empty interior
(or equivalently, dense complement).
Example 2: Sets of continuity and discontinuity. On X = R, consider the function
f defined by
n
(7.1)
f (x) = 1/q if x = p/q in lowest terms
0
if x is irrational
It is easy to see that f is continuous at the irrationals (why?) and discontinuous at the
rationals (why?). Does there exist a function with the reverse property, i.e. one that is
continuous at the rationals and discontinuous at the irrationals?
After spending an hour or two trying to construct such a function, one begins to suspect
that no such function can exist. But how to prove it? Once again, the Baire category
theorem comes to the rescue.
In order to measure quantitatively the continuity or discontinuity of a function, we introduce the concept of oscillation. Suppose that f is a map from a metric space X into another
metric space Y . For any subset S X, we define the oscillation of f on S as
f (S) = diam(f [S]) = sup dY (f (x), f (x ))

(7.2)

x,x S

(its value is a nonnegative real number or +). Then, for any x X, we define the
oscillation of f at x as
f (x) = inf f (B(x, ))
(7.3)
>0

(its value is again a nonnegative real number or +). Clearly f is continuous at x0 if

and only if f (x0 ) = 0 (why?). When f is discontinuous at x0 , f (x0 ) gives a quantitative
measure of the size of the discontinuity.

Note that if f (x0 ) < c (where c is some real number), then f (x) < c for all x in a
sufficiently small neighborhood of x0 (why?).5 So the set {x: f (x) < c} is open. But then
the set C of points of continuity of f can be written as
C =

{x: f (x) < 1/n}

(7.4)

n=1

and hence is a G . We have therefore proven:

Theorem 7.7 Let f be any map from a metric space X into another metric space Y . Then
the set of points of continuity of f is a G , and the set of points of discontinuity is an F .
Putting this together with the result of Example 1 that the rationals are not a G subset
of R, we conclude that there does not exist a map from R into R (or indeed into any metric
space) that is continuous at the rationals and discontinuous at the irrationals.
We also have the following corollary showing that the set of points of discontinuity of a
function f is either small (i.e. meager) or else somewhat large (i.e. contains a nonempty
open set):
Corollary 7.8 Let f be any map from a metric space X into another metric space Y . Then
the set of points of discontinuity of f is either meager or else has nonempty interior.
Proof. This follows immediately from Theorem 7.7 and Proposition 7.6. 
Remarks. 1. The nonexistence of a function that is continuous at the rationals and
discontinuous at the irrationals can alternatively be deduced from Corollary 7.8, by observing
that the irrationals have empty interior but are not meager.
2. Theorem 7.7 and Corollary 7.8 hold whether or not X is complete. The remarks given
after Proposition 7.6 apply also here. 
Example 3: Nowhere-differentiable continuous functions. Most of the continuous functions studied in elementary calculus are differentiable everywhere or at worst,
differentiable everywhere except at some finite set of points (e.g. the function |x| is nondifferentiable at x = 0). So it was something of a surprise when Weierstrass published in 1872
his famous example of a function that is continuous everywhere and differentiable nowhere.6
5

In other words, f is an upper semicontinuous function on X with values in [0, +].

Weierstrass example was

f (x) =

an cos(bn x)

n=0

where 0 < a < 1, b is an odd positive integer, and ab > 1 + (3/2). The condition 0 < a < 1 guarantees that
the series is uniformly convergent and hence that f is continuous; the remaining two conditions on b are used
by Weierstrass to prove that f is nowhere dierentiable. But these limitations on the allowed values of b
look a bit unnatural, since intuitively one expects that taking b to be any real number with ab > 1 should
suce to make f nowhere dierentiable, because the formal series for f (x) then looks likely to diverge for
all x (or maybe only for most x? that is the trouble). This intuitive guess does turn out to be correct, but it
took over four decades after Weierstrass work for this to be proven. Finally, in 1916 the celebrated English
mathematician G.H. Hardy (18771947) proved that f is nowhere dierentiable whenever ab 1.

But applying the Baire category theorem provides an even bigger surprise: it turns out that
most continuous functions are nowhere differentiable! Here, of course, most has to be
understood in the sense of Baire category: the precise claim is that the functions that are
differentiable at at least one point form a meager set in the space C[0, 1] of continuous functions. Indeed, the functions that have a finite one-sided derivative at at least one point, or
even a bounded difference quotient on at least one side at at least one point, form a meager
set in C[0, 1]. Let us prove this, as follows:
First let us be precise about what we are asserting. We say that a function f C[0, 1]
has a bounded right difference quotient at x [0, 1) in case
 f (x + h) f (x) 


lim sup 
 < ,
h
h0

or in other words

(7.5)

 f (x + h) f (x) 


there exists n < and > 0 such that 
 n for all h (0, ] .
h

(7.6)

So, for x [0, 1), h (0, 1 x] and n N, let us define

Ax,h,n

 f (x + h) f (x) 
o
n


= f C[0, 1]: 
n .
h

(7.7)

This is the set of continuous functions on [0, 1] whose difference quotient at x with step +h
is bounded in absolute value by n. The set of continuous functions on [0, 1] that have a
bounded right difference quotient at at least one point is therefore
A =

Ax,h,n

(7.8a)

n=1 m=1 0x<1 0<hmin(1x,1/m)

Ax,h,n

(7.8b)

n=1 0x11/n 0<h1/n

(why is the second form equivalent to the first?). So let us define

[
\
An =
Ax,h,n .

(7.9)

0x11/n 0<h1/n

We will prove that An is closed and nowhere dense, and hence that A =

An is meager.

n=1

Proof that An is closed. Consider a sequence (fk ) in An converging (in sup norm)
to f C[0, 1]. For each k there exists a point xk [0, 1 1/n] such that
|fk (xk + h) fk (xk )| nh for all h (0, 1/n] .

(7.10)

Then, by compactness, there exists a subsequence of (xk ) converging to some point x

[0, 1 1/n]; so, replacing (fk ) and (xk ) by this subsequence, we may assume for simplicity
8

of notation that the original sequence (xk ) tends to x. Taking k in (7.10), we obtain7
|f (x + h) f (x)| nh for all h (0, 1/n] ,

(7.11)

which shows that f An .

Proof that An has empty interior. Why does An not contain any open ball? The
reason is that near (in sup norm) to any continuous function there is another continuous
function whose slope is very badly behaved. To prove this, we shall first observe that near
any continuous function there is another continuous function whose slope is well behaved,
and then modify this latter function so that the slope is badly behaved. (This may seem a
rather contorted way of doing things, but you will see why it is necessary.)
So, the first step is to observe that the Lipschitz functions i.e. those for which there
exists M < such that |g(x) g(y)| M|x y| for all x, y [0, 1] are dense in
C[0, 1]. This follows immediately from the Weierstrass approximation theorem, because the
polynomials are obviously Lipschitz (why?); or it follows in a more elementary way by using
the fact that every continuous function on [0, 1] is uniformly continuous, so we can uniformly
approximate it by a piecewise linear function (you should supply the details of this proof).
Either way, we conclude that given any f C[0, 1] and any > 0, we can find g C[0, 1]
and M < such that kf gk < and |g(x) g(y)| M|x y| for all x, y [0, 1].
Now, for each positive integer N, let TN C[0, 1] be a triangular wave of amplitude 1
and half-period 1/N, i.e. the function that takes the values TN (j/N) = (1)j for j integer
and that is defined to be linear on each interval [j/N, (j + 1)/N]. On each such interval, TN
has slope 2N; so given any point x [0, 1) we have
 T (x + h) T (x) 
 N

N

 = 2N
h

(7.12)

for all sufficiently small h > 0 (why? can you see what determines how small h has to
be?). It follows that the function g + TN has right difference quotient at least 2N M in
magnitude at every point of [0, 1) (why?). So if we choose N > (M + n)/(2), we conclude
that g + TN
/ An . Since kf (g + TN )k < 2, we have proven that An does not contain
the open ball B(f, 2). But since f C[0, 1] and > 0 were arbitrary, we have proven that
An does not contain any open ball, i.e. An has empty interior.

S
This completes the proof that A =
An is meager.
n=1

Of course, a completely analogous proof works for left rather than right difference quotients. Since the union of two meager sets is meager, we conclude that the functions that
have a bounded difference quotient on at least one side at at least one point form a meager
set in C[0, 1]. 
7

Here we use the following simple lemma: If fk converges uniformly (i.e. in sup norm) to f , and xk
converges to x, then fk (xk ) converges to f (x). This follows from
|f (x) fk (xk )|

|f (x) f (xk )| + |f (xk ) fk (xk )|

|f (x) f (xk )| + kf fk k

and the continuity of the limiting function f .

Several other interesting applications of the Baire category theorem can be found in
Giles, Introduction to the Analysis of Normed Linear Spaces, Section 9; and many fascinating
applications can be found in Oxtoby, Measure and Category.

The uniform boundedness theorem

Let X and Y be normed linear spaces, and let F B(X, Y ) be a family of bounded
(i.e. continuous) linear maps from X to Y . We say that the family F is bounded (or
uniformly bounded) if it is bounded as a subset of the normed linear space B(X, Y ),
i.e. if {kT kXY : T F } is a bounded set of real numbers. We say that the family F is
pointwise bounded if, for each x X, the set {T x: T F } is bounded as a subset of the
normed linear space Y , i.e. if {kT xkY : T F } is a bounded set of real numbers. Clearly a
bounded family is pointwise bounded (why?). The rather surprising fact is that the converse
is also true, provided that X is complete. This important result is known as the uniform
boundedness theorem (or principle of uniform boundedness or BanachSteinhaus
theorem):8
Theorem 7.9 (uniform boundedness theorem) Let X be a Banach space and Y a normed
linear space, and let F be a family of bounded linear maps from X to Y . If F is pointwise
bounded, then it is bounded.
The key step in the proof of the uniform boundedness theorem is the following lemma,
which is based on the Baire category theorem. We say that a subset A of a normed linear
space is symmetric if x A implies x A.
Lemma 7.10 Let X be a Banach space, and let C be a closed convex symmetric subset of

S
X satisfying
nC = X. Then C is a neighborhood of 0 [that is, C contains an open ball
n=1

B(0, ) for some > 0].

Proof. By the weak form of the Baire category theorem (Corollary 7.4b ), one of the sets nC
must have nonempty interior, hence C must have nonempty interior (why?). So C contains
some open ball B(x, ) with > 0. Since C is symmetric, C also contains the open ball
B(x, ). And then, since C is convex, C also contains the open ball B(0, ) (why?). 
Remark. Lemma 7.10 actually holds without the hypothesis that C is symmetric
(though we will not need this fact). To see this, it suffices to apply Lemma 7.10 to the set

S
D = C (C). Clearly D is closed, convex and symmetric (why?). To see that
nD = X,
n=1

observe first that since 0 C (why?) and C is convex, for each x X there exists an integer
nx such that x nC for all n nx (why?); so if n max(nx , nx ) we have x nD. 
8

The uniform boundedness theorem was rst proven in 1922, independently by the Austrian mathematician Hans Hahn (18791934) and the Polish mathematician Stefan Banach (18921945). The proof using the
Baire category theorem was published in 1927 by Banach together with his Polish colleague Hugo Steinhaus
(18871972); the idea had been suggested by another Polish mathematician, Stanislaw Saks (18971942).

10

Proof of the uniform boundedness theorem. Let

C = {x X: kT xkY 1 for all T F } .

(7.13)

C is closed because each T F is continuous (how is the for all handled?). C is symmetric
(why?) and convex (why?). Finally, the pointwise boundedness of F tells us that for each
x X there exists an integer n (depending on x) such that kT xkY n for all T F , or in

S
other words x nC. This means that
nC = X. Lemma 7.10 then tells us that C contains
n=1

an open ball B(0, ) for some > 0. It follows that kT xkY 1 for all T F whenever
kxkX < [and hence also whenever kxkX why?], or in other words kT xkY 1 kxkX
for all T F and all x X, or in other words kT kXY 1 for all T F . This shows
that the family F is bounded. 

Example. Here is a simple example showing that the conclusion of the uniform boundedness theorem can fail if X is incomplete: Let X be the space c00 of sequences x = (x1 , x2 , . . .)
with at most finitely many nonzero entries, equipped with the sup norm (i.e. considered as
a linear subspace of ), and take Y = R. For n = 1, 2, . . ., let n X be defined by
n (x) = nxn . Then for any x X there is an integer Nx such that xn = 0 for n > Nx , so we
have
|n (x)| Nx kxk for all n .
(7.14)
This shows that the family {n }
n=1 is pointwise bounded. But
kn k = n for each n

(7.15)

(why?), so the family {n }

n=1 is not bounded. 
Remark. In the proof of Lemma 7.10 and hence the uniform boundedness theorem, we
did not really use as such the assumption that X is complete; all we used is the fact that
it is nonmeager in itself (where did we use this?). So the uniform boundedness theorem
holds under this weaker hypothesis. There do exist incomplete normed linear spaces that
are nonmeager in themselves, so this extra generality is not vacuous; but such spaces rarely
arise in applications. Please observe that the example space c00 is indeed meager in itself
(why?). 
The standard textbook proof of the uniform boundedness theorem is the one I have
just presented, using the Baire category theorem. As you have just seen, this proof is
quite simple; but its reliance on the Baire category theorem makes it not completely
elementary.
By contrast, the original proofs of the uniform boundedness theorem given by Hans
Hahn and Stefan Banach in 1922 were quite dierent: they began from the assumption
that sup kT k = and used a gliding hump technique to construct a sequence (Tn )
T F

in F and a point x X such that lim kTn xk = . These proofs are elementary, but
n

11

the details are a bit ddly. Here is a really simple proof along similar lines, which I
discovered three years ago while preparing this course9 :
Lemma. Let T be a bounded linear operator from a normed linear space X to a
normed linear space Y . Then for any x X and r > 0, we have
sup
x B(x,r)

kT x k kT kr .

(7.16)

Proof. For X we have

max{kT (x + )k, kT (x )k}

1
2 [kT (x

+ )k + kT (x )k] kT k ,

(7.17)

where the second uses the triangle inequality in the form k k kk + kk. Now
take the supremum over B(0, r). 
Proof of the uniform boundedness theorem. Suppose that supT F kT k = ,
n
and choose (Tn )
n=1 in F such that kTn k 4 . Then set x0 = 0, and for n 1 use
the lemma to choose inductively xn X such that kxn xn1 k 3n and kTn xn k
2 n
kTn k. The sequence (xn ) is Cauchy, hence convergent to some x X; and it is
33
easy to see that kx xn k 12 3n and hence kTn xk 61 3n kTn k 61 (4/3)n . 

One important application of the uniform boundedness theorem concerns pointwise convergent sequences of linear mappings. So let X and Y again be normed linear spaces, and let
(Tn ) be a sequence of continuous linear maps from X to Y . We say that the sequence (Tn )
is pointwise convergent to a mapping T if, for each x X, Tn x is convergent to T x in Y .
In such a situation the limiting mapping T is obviously linear, but need it be continuous?
In general the answer is no:
Example. Consider once again X = c00 , equipped with the sup norm, and Y = R. This
time let n X be defined by
n
X
xi .
(7.18)
n (x) =
i=1

Then the sequence (n ) is pointwise convergent to the linear mapping defined by

(x) =

xi

(7.19)

i=1

(why is well-defined? why is (n ) pointwise convergent to ?). But is unbounded,

because the vector x(n) = (1, 1, . . . , 1, 0, 0, . . .) consisting of n ones followed by zeros has
kx(n) k = 1 but (x(n) ) = n. 
But once again, you will observe in this example that X is incomplete (and indeed meager
in itself). If X is complete (or more generally if X is nonmeager in itself), then the limiting
mapping T does have to be continuous:
9

See A.D. Sokal, A really simple elementary proof of the uniform boundedness theorem, Amer. Math.
Monthly 118, 450452 (2011), also available at http://arxiv.org/abs/1005.1585.

12

Corollary 7.11 Let X be a Banach space and Y a normed linear space, and let (Tn ) be a
sequence of continuous linear maps from X to Y that is pointwise convergent to a mapping
T . Then T is continuous.
Proof. Since the sequence (Tn ) is pointwise convergent, it is pointwise bounded. Therefore,
by the uniform boundedness theorem, the sequence (Tn ) is bounded, i.e. there exists M <
such that kTn kXY M for all n. This can be rephrased as saying that kTn xkY MkxkX
for all n and all x X (why?). Taking n , we have Tn x T x in Y for all x X
(why?), hence kT xk MkxkX for all x X (why?). This proves that T is bounded, i.e.
continuous. 
Here is another important application of the uniform boundedness theorem. Let X, Y, Z
be metric spaces. Then a map f : X Y Z is said to be separately continuous if f (x, y)
is continuous as a function of x for each fixed value of y, and vice versa; or in more detail
xn x, y Y = f (xn , y) f (x, y)
yn y, x X = f (x, yn ) f (x, y)
The map f is said to be jointly continuous if it is continuous as a map from the product
metric space X Y (with any of its standard equivalent metrics) to Z; or in more detail
xn x, yn y = f (xn , yn ) f (x, y)
Clearly joint continuity implies separate continuity (why?), but is the reverse true? Cauchy
wrote in his 1821 Cours dAnalyse that a separately continuous function of two real variables
is jointly continuous, but this is false! (Even great mathematicians can make mistakes.) You
perhaps saw in your real analysis course the following counterexample10 :
( xy
if (x, y) 6= (0, 0)
f (x, y) =
(7.20)
x2 + y 2
0
if (x, y) = (0, 0)
(Why is this function separately continuous? Why is it not jointly continuous?)
So a separately continuous function of two variables need not be jointly continuous. It is
therefore somewhat surprising that a separately continuous bilinear mapping defined on a
pair of Banach spaces is always jointly continuous:
Proposition 7.12 (joint continuity of separately continuous bilinear mappings)
Let X, Y, Z be normed linear spaces, with either X or Y (or both) complete, and let T : X
Y Z be a separately continuous bilinear mapping. Then T is jointly continuous.
10

Due to the German mathematician Hermann Schwarz (18431921) in 1872. A similar but slightly more
complicated example was given two years earlier by the German mathematician Johannes Thomae (1840
1921), crediting his colleague Eduard Heine (18211881).

13

Proof. Suppose that X is complete. For each x X, consider the linear mapping
T (x, ): Y Z. This is a bounded linear mapping (why?), so there exists a constant
Kx < such that
kT (x, y)kZ Kx kykY for all y Y .
(7.21)
Next consider, for each y Y with kykY 1, the linear mapping T ( , y): X Z. This
is a bounded linear mapping (why?), and it follows from (7.21) that the collection of these
mappings, {T ( , y): kykY 1}, is pointwise bounded on X (why?). Since X is complete,
the uniform boundedness theorem tells us that this collection of mappings is bounded, i.e.
there exists a constant K < such that
kT ( , y)kXZ K

for all y Y with kykY 1 ,

(7.22)

or equivalently
kT (x, y)kZ KkxkX

for all x X and y Y with kykY 1 ,

(7.23)

or equivalently
kT (x, y)kZ KkxkX kykY

for all x X and y Y .

(7.24)

This shows that T is jointly continuous. 

Example. Here is a simple example showing that we need at least one of the spaces X
and Y to be complete, otherwise the conclusion can fail: Let X = Y = c00 equipped with the

P
sup norm, and Z = R; then let T (x, y) =
xn yn . Now T is well-defined (why?); and it is
n=1

separately continuous, because kT (x, )kY Z = kxk1 < (why?) and likewise for T ( , y).
But if x(n) = (1, 1, . . . , 1, 0, 0, . . .) is the vector consisting of n
ones followed by zeros, we
(n)
(n)
(n)
(n)
have kx k =
1 but T (x , x ) = n; so the sequence (x / n) converges to zero but
T (x(n) / n, x(n) / n) = 1 does not converge to zero; so T is not jointly continuous.
What would happen in this example if we took one of the two domain spaces (say, X)
to be rather than c00 ? Then we would satisfy the hypothesis of Proposition 7.12 that
at least one of the domain spaces be complete, and T would remain well-defined (why?),
but T would fail to be separately continuous! Indeed, T (x, ) would be an unbounded linear
functional on Y = c00 whenever x \ 1 (why?). 
The uniform boundedness theorem and its corollaries are extremely useful in real and
functional analysis. In Problem Set #7 I will give you some typical applications.

The open mapping theorem

Let X and Y be metric spaces, and let f be a map from X to Y . We have frequently
used the elementary fact that f is continuous if and only if the inverse image of every open
set is open, i.e. U open in Y implies f 1 [U] open in X. Let us now say that the map f is an
open mapping if the direct image of every open set is open, i.e. U open in X implies f [U]
open in Y .
14

A continuous mapping need not be open, and an open mapping need not be continuous.
For instance, if X is the real line with the usual metric, Y is the real line with the discrete
metric, and f : X Y is the identity map, then f is open but not continuous (why?), and
f 1 is continuous but not open (why?).11 However, if f : X Y is a bijection, then f 1 is
continuous if and only if f is open (why?). This is the major reason that we are interested
in open mappings: when we encounter a bijective map that is continuous, it is natural to
ask whether its inverse is continuous as well. In particular we now want to pose questions
of this type for linear operators from one Banach space to another.
You will recall (Example 6 of Handout #3) the following simple example of a bounded (i.e.
continuous) linear map with an unbounded (i.e. discontinuous) inverse: Consider T : 2 2
defined by
1
(T x)j = xj .
(7.25)
j
It is easy to see that T is bounded (of operator norm kT k2 2 = 1). Furthermore, T is
an injection (why?). But T is not a surjection; rather, its image is the proper dense linear
subspace

X
k 2 |xk |2 < }
(7.26)
M = {x 2 :
k=1

(why?). Since T is a linear bijection from to M, the inverse map T 1 is well-defined as a

map from M to 2 ; it is obviously given by
(T 1 x)j = j xj ,

(7.27)

which is unbounded (why?). So a bounded linear bijection T : 2 M can have an unbounded

inverse T 1 : M 2 .
As remarked already in Handout #3, this pathology arises from the fact that the range
space M is not complete. In fact, we have the following amazing results, due principally to
Banach in the late 1920s:
Theorem 7.13 (open mapping theorem) A bounded linear surjection of a Banach space
X onto a Banach space Y is an open mapping.
Corollary 7.14 (inverse mapping theorem) A bounded linear bijection of a Banach space
X onto a Banach space Y has a bounded inverse (and hence is a topological isomorphism).
Do you see why the inverse mapping theorem is an immediate consequence of the open
mapping theorem?
We will actually prove the following strong version of the open mapping theorem:
Theorem 7.15 (open mapping theorem, strong version) Let T be a bounded linear
map from a Banach space X into a normed linear space Y . If the image T [X] is nonmeager
in Y , then T is surjective (i.e. T [X] = Y ) and an open mapping.
11

The fact that a continuous mapping need not be open was remarked already in Handout #1, using the
following example: the map f (x) = x2 is continuous from R to R, but the image of the open set (1, 1) is
the non-open set [0, 1).

15

Do you see why the usual form of the open mapping theorem follows immediately from
this version together with the (weak form of the) Baire category theorem?
Remark. With a bit more work one can prove not only that T is surjective and open,
but also that Y is complete and topologically isomorphic to the quotient space X/(ker T )
under the map Te: X/(ker T ) Y given by Te(x + ker T ) = T (x). 
We will prove the strong version of the open mapping theorem by a series of lemmas.
Let us denote by B (resp. B ) the open unit ball in X (resp. Y ).
The first lemma is almost trivial, but worth stating explicitly:

Lemma 7.16 Let X and Y be normed linear spaces, and let T : X Y be a linear mapping
(not necessarily even continuous!). Then the following are equivalent:
(a) T is an open mapping.
(b) T [B] is an open neighborhood of 0 in Y .
(c) T [B] is a neighborhood of 0 in Y .
Proof. If T is an open mapping, then clearly T [B] is an open neighborhood of 0 in Y
(why?). Conversely, suppose that T [B] is a neighborhood of 0 in Y , and let U X be an
open set. Given any y T [U], choose x U such that y = T x; then there exists > 0 such
that B(x, ) U (why?). But it follows from this that y + T [B] T [U], which proves that
T [U] is open (why?). 
So, to prove that T is an open mapping, we will prove that T [B] is a neighborhood of 0
in Y . We begin with a lemma asserting something slightly weaker, namely that the closure
of T [B] is a neighborhood of 0 in Y :
Lemma 7.17 Let X and Y be normed linear spaces, and let T : X Y be a linear mapping
(not necessarily even continuous!). If T [X] is nonmeager in Y , then T [B] is a neighborhood
of 0 in Y .
Proof. We have T [X] =

nT [B]. Since T [X] is nonmeager, the sets nT [B] cannot all be

n=1

nowhere dense, hence T [B] cannot be nowhere dense (why?), or in other words T [B] must
have nonempty interior. Since the set T [B] is convex (why?) and symmetric (why?), the
argument used in the proof of Lemma 7.10 tells us that T [B] is a neighborhood of 0 in Y .

And finally, we show how to infer from this weaker statement that T [B] is itself a neighborhood of 0 in Y :
Lemma 7.18 Let X be a Banach space and Y a normed linear space, and let T : X Y
be a continuous linear mapping. If T [B] is a neighborhood of 0 in Y , then T [B] is also a
neighborhood of 0 in Y .
16

Proof. Suppose that T [B] is a neighborhood of 0 in Y , so that there exists > 0 such
that B T [B]. Then for y B we have y T [B], so we can choose y1 T [B] such
that ky y1 k < /2 (why?); and we can choose x1 B such that T x1 = y1 (why?). Then
y y1 (/2)B 12 T [B], so we can choose y2 21 T [B] such that k(y y1 ) y2 k < /4
(why?); and we can choose x2 12 B such that T x2 = y2 . Continuing inductively, we obtain
sequences (xn ) in X and (yn ) in Y such that xn 2(n1) B, yn = T xn and
ky (y1 + y2 + . . . + yn )k < /2n .
Now write sn =

n
P

k=1

(7.28)

xk . Since kxn k < 1/2n1 , we have

ksm sn k

m
X

k=n+1

kxk k <

1
2n1

whenever m > n ,

(7.29)

hence (sn ) is a Cauchy sequence in X. Since X is complete, (sn ) is convergent to some

P
kxk k < 2. Since ky T sn k < /2n (why?) and
point s X. Moreover, we have ksk
k=1

T sn T s (why? what hypothesis about T is being used here?), we have y = T s. This shows
that y 2T [B]. Since y was an arbitrary element of B , we have shown that B 2T [B],
i.e. (/2)B T [B]. Hence T [B] is a neighborhood of 0 in Y . 

Proof of the strong version of the open mapping theorem. Let T be a bounded
linear map from a Banach space X into a normed linear space Y . If T [X] is nonmeager in Y ,
then Lemma 7.17 implies that T [B] is a neighborhood of 0 in Y . But then Lemma 7.18 implies
that T [B] is a neighborhood of 0 in Y . And this means, by Lemma 7.16, that T is an open
mapping. And a linear map that is an open mapping is obviously surjective (why?). 

The closed graph theorem

Let X and Y be metric spaces. Then a function f : X Y is said to have a closed
graph if
graph(f ) = {(x, f (x)): x X}
(7.30)
is a closed subset of the product space X Y (equipped with any one of its standard metrics).
Note in particular that if f is bijective, then f has a closed graph if and only if f 1 does
(why?).
How does having a closed graph compare to the related property of continuity? Consider
the following three statements concerning a sequence (xn ) in X and elements x X, y Y :
(a) xn x
(b) f (xn ) y
(c) y = f (x)
17

Continuity is the statement that (a) implies (b) and (c) [for the appropriately chosen y].
Having a closed graph is the statement that (a) and (b) together imply (c).
So a continuous function always has a closed graph, but the converse is false: for instance,
the function f : R R defined by
n
f (x) = 1/x if x 6= 0
(7.31)
a
if x = 0

(for any chosen a R) has a closed graph (why?) but is not continuous (why?).
It is therefore somewhat surprising that for linear operators from one Banach space into
another, having a closed graph automatically implies continuity:
Theorem 7.19 (closed graph theorem) Let X and Y be Banach spaces, and let T : X
Y be a linear map. Then T is continuous if and only if it has a closed graph.

Proof. We have already proved that every continuous map has a closed graph, so let us
prove the converse. This will be a fairly simple application of the open mapping theorem.
Since X and Y are Banach spaces, the product space X Y is a Banach space when
equipped with the norm k(x, y)kXY = kxkX + kykY . Since graph(T ) is a closed linear
subspace of X Y , it is a Banach space in its own right. Now consider the projection operator
1 : graph(T ) X defined by 1 (x, y) = x. Clearly 1 is a bijection of graph(T ) onto X
(why?). Moreover, since k1 (x, y)kX = kxkX k(x, y)kXY , the map 1 is continuous (of
norm 1). The open mapping theorem then implies that 1
1 is a continuous linear map of X
into (in fact onto) graph(T ). On the other hand, the projection operator 2 : graph(T ) Y
defined by 2 (x, y) = y is also continuous (of norm 1) since k2 (x, y)kX = kykY
k(x, y)kXY . It follows that T = 2 1
1 is a continuous map of X into Y . 
The completeness of both X and Y is crucial in the closed graph theorem; otherwise the
conclusion can fail:
Example 1: An unbounded multiplication operator. Suppose that we try to define
an unbounded linear operator T : 1 1 by
(T xk ) = kxk .

(7.32)

The trouble is that this formula does not map 1 into itself; rather, we must restrict attention
to the proper dense linear subspace M ( 1 defined by
1

M = {x :

X
k=1

k|xk | < } .

(7.33)

(Why is M dense in 1 ?) Then T : M 1 is a well-defined linear map. Furthermore, T has

a closed graph: for if (x(n) ) is a sequence in M converging in 1 norm to x M, and we also
know that (T x(n) ) converges in 1 norm to some y 1 , then it follows that T x = y (why?).
[Alternatively, it suffices to observe that T is a bijection from M onto 1 (why?), and that
the inverse map T 1 is bounded (why?) and hence has a closed graph; so T must have a
18

closed graph as well.] But T is not continuous. This does not contradict the closed graph
theorem, because the domain space M is not complete.
This example could equally well be carried out on any of the spaces p with 1 p < ,
or in c0 . (What changes if we try it in ?)
Example 2: A differential operator. In the space C[0, 1], let us try to define the
operator T of differentiation. The trouble (as discussed already in Example 5 of Handout #3)
is that differentiation makes no sense for arbitrary functions f C[0, 1] (at least, not if we
want the derivative f to also belong to C[0, 1]). Rather, we need to restrict ourselves to (for
example) the space C 1 [0, 1] consisting of functions that are once continuously differentiable
on [0, 1] (where we insist on having also one-sided derivatives at the two endpoints, and the
derivative f is supposed to be continuous on all of [0, 1], including at the endpoints). This
is a proper dense linear subspace of C[0, 1] (why is it dense? why is it not closed?). Then the
map T : f 7 f is a well-defined linear map from C 1 [0, 1] into C[0, 1]. Furthermore, T has
a closed graph: for if (fn ) is a sequence in C 1 [0, 1] converging in sup norm to f C 1 [0, 1],
and we also know that (fn ) converges in sup norm to some g C[0, 1], then it follows that
f = g (why?). On the other hand, T is not continuous (why?). Once again, this does not
contradict the closed graph theorem, because the domain space C 1 [0, 1] with the sup norm
is not complete.
Example 3. A more complicated example shows that the completeness of Y is also
essential in the closed graph theorem: see e.g. John B. Conway, A Course in Functional
Analysis, p. 92.
Some final remarks. In this course we have concentrated our attention on bounded
(i.e. continuous) linear operators; but unbounded linear operators do play a central role in
many applications of functional analysis. For instance, differential operators are invariably
unbounded (as Example 2 makes clear) whenever they are considered as acting on the usual
normed function spaces such as C(X) or Lp (X); so applications of functional analysis to the
study of partial differential equations (PDE) cannot avoid dealing with unbounded operators. Likewise, most of the operators arising in quantum mechanics (position, momentum,
Hamiltonians, . . . ) are unbounded operators on a Hilbert space.
What the closed graph theorem shows is that unbounded linear operators cannot be
defined in any sensible way (e.g. any way that gives them a closed graph) on the whole of
a Banach space. Rather, as Examples 1 and 2 make clear, unbounded linear operators are
most naturally defined on a domain D that is a proper dense subspace of a Banach space.
To address this situation, it makes sense to generalize slightly the concept of a closed
graph, as follows: Let X and Y be metric spaces, and let D be a subset of X (the domain).
Then a function f : D Y is said to have a closed graph (relative to X and Y ) if
graph(f ) = {(x, f (x)): x D}

(7.34)

is a closed subset of the product space X Y .

Please note that we are now considering the domain D as a subset of a possibly bigger
space X, and that whether f has a closed graph or not can depend on the choice of X. For
19

instance, the identity map f (x) = x with D = (0, 1) has a closed graph if we take X = (0, 1)
and Y (0, 1), or Y = (0, 1) and X (0, 1), but not if we take e.g. X = Y = [0, 1].
In this new context, we can once again rephrase the closed-graph condition in terms of
sequences: a map f has a closed graph (in the new sense) if and only if the situation
xn D,

xn x and f (xn ) y

(7.35)

implies that
xD

and f (x) = y .

(7.36)

The trouble is that continuity no longer automatically implies the closed-graph property (as
the preceding example shows). But it is fairly easy to prove the following facts (I leave the
proofs as an exercise for you):
Proposition 7.20 Let X and Y be metric spaces, and let D be a closed subset of X. Then
a continuous function f : D Y has a closed graph.
Proposition 7.21 Let X and Y be normed linear spaces, with Y complete, and let D be a
linear subspace of X. Then a continuous linear map T : D Y has a closed graph if and
only if D is closed.
The important point is now that the typical unbounded operators, such as those in
Examples 1 and 2, are closed also in the new sense (which is stronger than the old sense):
Example 1, revisited. We previously considered T as an operator from M to 1 , and we
observed that it has a closed graph in the sense that if (x(n) ) is a sequence in M converging
in 1 norm to x M, and (T x(n) ) converges in 1 norm to some y 1 , then it follows that
T x = y. But let us now consider T as an operator with domain D = M lying inside the
space X = 1 . Then, to verify that T has a closed graph, we need to prove the stronger
statement that if (x(n) ) is a sequence in M converging in 1 norm to x 1 (not necessarily
in M!), and (T x(n) ) converges in 1 norm to some y 1 , then it follows that x M and
T x = y. Do you see why this is true?
Example 2, revisited. We previously considered T as an operator from C 1 [0, 1] to
C[0, 1], and we observed that it has a closed graph in the sense that if (fn ) is a sequence
in C 1 [0, 1] converging in sup norm to f C 1 [0, 1], and (fn ) converges in sup norm to some
g C[0, 1], then it follows that f = g. But let us now consider T as an operator with domain
D = C 1 [0, 1] lying inside the space X = C[0, 1]. Then, to verify that T has a closed graph,
we need to prove the stronger statement that if (fn ) is a sequence in C 1 [0, 1] converging in
sup norm to f C[0, 1] (not necessarily in C 1 [0, 1]!), and (fn ) converges in sup norm to some
g C[0, 1], then it follows that f C 1 [0, 1] and f = g. Do you see why this is true?
Many (but not all) of the results concerning bounded linear operators can be carried
over, with some modifications and some extra work, to linear operators having a closed
graph (usually called closed linear operators for short). But that topic belongs to a more
advanced course in functional analysis.
20